基于BP神经网络的二元混合液体自燃温度预测

为了给工业界提供一种快速预测二元混合液体自燃温度的有效途径,将试验所测不同组分及配比的168个二元混合液体的自燃温度作为期望输出,将基于电性拓扑状态指数(ETSI)理论、引入混合ETSI概念而计算出的9种原子类型所对应的混合ETSI作为输入,采用三层BP神经网络技术建立了根据原子类型混合ETSI来预测混合液体自燃温度的BP神经网络模型,并应用改进的Garson算法进行多参数敏感性分析。经模型评价验证及稳定性分析,得到训练集的决定系数R2为0.965,平均绝对误差MAE为11.892 K,测试集的交叉验证系数Q2ext为0.923,平均绝对误差MAE为15.530 K,发现该模型的预测性能优于已有的多元非线性回归(MNR)模型,表明BP神经网络模型具有较好的拟合能力和预测能力,对烷、醇类混合体系自燃温度的预测精度最佳。

Forecasting the auto-ignition temperatures of binary liquid mixtures based on BP neural network

The present paper intends to provide a quick effective way for forecasting auto-ignition temperatures (AIT) of the binary flammable liquid mixtures for the chemical engineering projects,the AIT of 168 binary flammable liquid mixtures composed of different components and volume ratios to be chosen as the expected outputs.It has also aimed at selecting the mixed ETSI values of 9 kinds of atom types worked out based on the electro-topological state indices (ETSI) theory as the input variables.It is just for the above said purposes that we have developed a three-layer backward-propagating neural network (BPNN) model for predicting the AIT of binary liquid mixtures according to the atom-type mixed ETSI values and divide the dataset of 168 mixtures randomly into two categories,that is,the training set (140) and the testing set (28).At the same time,we have also selected the gradient descent with the momentum and the adaptive learning rate algorithm as the training function to avoid the slow convergence speed and the low learning accuracy.Thus,we have gained the optimal condition of the neural network by adjusting the various parameters via the trial-and-error method,and finding the final optimum structure of BP neural network equal to be 9-8-1.Furthermore,we have also applied the improved Garson algorithm to the multi-parameter sensitivity analysis in assessing the relative and absolute contributions of each atom type through the connection weights of the well trained BPNN model.And,then,we have done the evaluation validation and the stability analysis to validate the model we have proposed,which proves to be obviously superior over the currently existing multiple nonlinear regression (MNR) inductions in terms of the model generalization performance and the prediction accuracy.Along the line,we have identified and determined that the coefficient of the determination (R2) and the mean absolute error (MAE) of the training set should be equal to 0.965 and 11.892 K,respectively.And,consequently,the coefficient of cross validation (Q2ext) and MAE of the test set should be equal to O.923 and 15.530 K,respectively.Moreover,the results of the calculation have shown that the BPNN model enjoys nice fitting ability and forecasting effect,with the predicted error being within th_e allowable range of AIT experimental error.